RoboPianist: Dexterous Piano Playing with Deep Reinforcement Learning

Kevin Zakka, Philipp Wu, Laura Smith, Nimrod Gileadi, Taylor Howell, Xue Bin Peng, Sumeet Singh, Yuval Tassa, Pete Florence, Andy Zeng, Pieter Abbeel

Keywords: high-dimensional control, bi-manual dexterity, piano playing, reinforcement learning

TL;DR: We propose a deep RL system, RoboPianist, which can successfully learn to play 150 piano pieces in simulation with 2 anthropomorphic dexterous hands.

Abstract: Developing robotic hands capable of replicating human-like dexterity has been a grand challenge in robotics. While recent works have made impressive progress in the last few years with learning-based approaches, the studied tasks such as in-hand reorientation, relocation and simple tool use fail to capture the breadth of skills exhibited by human hands. To push the boundaries of high-dimensional control, we explore the task of controlling bi-manual anthropomorphic robotic hands to play the piano, which is targeted at testing high spatial and temporal precision, coordination, and planning, all with an underactuated system frequently making-and-breaking contacts. We propose a deep RL system, RoboPianist, which can successfully learn a diverse repertoire of 150 piano pieces. The resulting policies exhibit highly dexterous behaviors, and produce visually and acoustically pleasing performances. Website featuring videos, code, and datasets is available at https://robopianist.github.io/.

Student First Author: yes

Instructions: I have read the instructions for authors (https://corl2023.org/instructions-for-authors/)